Algorithm 1
Label Subspace Construction
1: Input: initial segmentation and classification
results; |
2: Output: co-occurrence table; |
3: for each superpixel
in image do
|
4: Compute the label histogram
of
; |
5: Sort the histogram
in descending order; |
6: Save the label index vector in array
Lindex; |
7:
the label that most of the
pixels in
take; |
8: for
to M do
|
9:
element of
Lindex; |
10:
; |
11: end for
|
12: end for
|
Algorithm 2
Context Guided BP
1: Input: image
for classification and training
sample |
2: Output: classification result; |
3: Perform mean-shift segmentation; |
4: Perform SVM classification; |
5: Construct label co-occurrence table; |
6: while not converge do
|
7: Pass message
for all edge
; |
8: end while
|
9: Compute the belief of each node; |
10: Assign a label to each node that minimize its belief; |
Table 1.
Number of Samples in Washington, DC
Class Name | Training Sample | Test Sample |
---|
Water | 638 | 29,610 |
Grass | 1517 | 39,648 |
Roof | 3394 | 40,829 |
Trail | 542 | 1783 |
Road | 2086 | 24,397 |
Shadow | 722 | 2355 |
Tree | 1622 | 24,457 |
Table 2.
Number of Samples in Wuhan
Class Name | Training Sample | Test Sample |
---|
Water | 1393 | 4644 |
Tree | 2028 | 6761 |
Grass | 949 | 3166 |
Crop | 4181 | 13,939 |
Red roof | 1979 | 6598 |
Blue roof | 363 | 1211 |
Dark roof | 3109 | 10,366 |
Bare soil | 2701 | 9006 |
Road | 1399 | 4665 |
Table 3.
Number of Samples in Colorado
Class Name | Training Sample | Test Sample |
---|
Cement | 800 | 28,150 |
Tree | 800 | 58,863 |
Grass | 800 | 88,481 |
Building | 800 | 39,361 |
Road | 800 | 38,073 |
Table 4.
Comparison of the Classification Results on the Washington,
DC Image
| Accuracy (%) | | |
---|
Methods | Water | Grass | Roof | Trail | Road | Shadow | Tree | OA (%) | kappa (%) |
---|
CBP |
93.12
| 94.43 |
88.36
|
94.11
|
84.14
|
92.52
| 92.64 |
90.82
|
88.58
|
SBP | 93.04 |
94.55
| 88.30 |
94.11
| 83.80 | 84.89 |
93.06
| 90.74 | 88.48 |
SVM | 75.63 | 92.52 | 82.67 | 89.00 | 82.67 | 78.25 | 89.85 | 83.73 | 80.82 |
Table 5.
Comparison of the Classification Results on the Wuhan
Image
| Accuracy (%) | | |
---|
Methods | Water | Tree | Grass | Crop | Red roof | Blue roof | Dark roof | Bare soil | Road | OA (%) | kappa (%) |
---|
CBP |
99.76
|
97.60
| 82.63 | 85.04 | 91.32 |
92.32
| 87.14 |
87.41
|
89.58
| 89.35 | 87.59 |
SBP |
99.76
|
97.60
| 82.98 |
85.24
|
91.97
| 87.45 |
88.15
|
87.41
|
89.58
|
89.56
|
87.83
|
SVM | 99.18 | 91.05 |
91.82
| 76.81 | 86.84 | 92.15 | 80.89 | 83.49 | 88.6 | 84.92 | 82.53 |
Table 6.
Comparison of the Classification Results on the Colorado
Image
| Accuracy (%) | | |
---|
Methods | Cement | Tree | Grass | Roof | Road | OA (%) | kappa (%) |
---|
CBP | 75.80 |
73.51
|
70.06
|
86.07
|
94.23
|
77.68
|
73.27
|
SBP |
80.09
| 71.19 | 68.33 | 81.67 | 91.16 | 77.17 | 72.71 |
SVM | 77.14 | 71.81 | 63.53 | 79.76 | 88.41 | 73.14 | 65.00 |
Table 7.
Comparison of Computation Time (in ms)
Method | Washington, DC | Wuhan | Colorado |
---|
CBP |
34
|
42
|
9
|
SBP | 290 | 263 | 46 |
Table 8.
Comparison of CBP and BBP on the Washington, DC
image
| Accuracy (%) | | | |
---|
Methods | Water | Grass | Roof | Trail | Road | Shadow | Tree | OA (%) | kappa (%) | Times (ms) |
---|
CBP |
93.12
| 94.43 |
88.36
| 94.11 |
84.14
|
92.52
| 92.64 |
90.82
|
88.58
|
34
|
BBP | 83.65 |
94.69
| 87.89 |
94.50
| 83.81 | 91.64 |
93.32
| 89.09 | 86.47 | 270 |
Table 9.
Comparison of CBP and BBP on the Wuhan Image
| Accuracy (%) | | | |
---|
Methods | Water | Tree | Grass | Crop | Red Roof | Blue Roof | Dark roof | Bare soil | Road | OA (%) | kappa (%) | Times (ms) |
---|
CBP |
99.76
|
97.60
| 82.63 | 85.04 | 91.32 | 92.32 | 87.14 |
87.41
|
89.58
| 89.35 | 87.59 |
42
|
BBP |
99.76
|
97.60
|
82.98
|
85.23
|
92.30
|
93.23
|
88.15
|
87.41
|
89.58
|
89.71
|
88.01
| 159 |
Table 10.
Comparison of CBP and BBP on the Colorado Image
Methods | Accuracy (%) | OA
(%) | kappa (%) | Times (ms) |
---|
Cement | Tree | Grass | Roof | Road |
---|
CBP | 75.80 | 73.51 |
70.06
|
86.07
|
94.23
|
77.68
|
73.27
|
9
|
BBP |
76.59
|
73.67
| 68.77 | 85.50 | 94.21 | 77.23 | 72.28 | 35 |